Chai Discovery: Turning Drug Design Into a Scaling Problem

Key takeaways from the Sequoia Capital interview with Chai Discovery co-founders Joshua MeierDealroom has a profile for this one. Try Dealroom → and Matthew McPartlonDealroom has a profile for this one. Try Dealroom → (August 2026):

Drug discovery as a scaling problem. Chai frames itself as a foundation-model lab for biology, betting that molecular design responds to the same "bitter lesson" as the rest of machine learning — scaling data, models and compute — rather than being treated as a bespoke scientific field.

From discovery to design. The founders want to move biology from screening millions or billions of molecules for a needle in a haystack towards specifying the "dream state" of a molecule and having the model materialise it, making drug work look more like engineering with reusable abstraction layers.

Hit rates jumped from 0.1% to 15%. When the company started, state-of-the-art antibody design bound roughly one in a thousand molecules; its Chai-2 model reached about a 15% success rate, so screening a thousand molecules now returns around 150 binders — enough to gather meaningful statistics on drug-like properties.

Simplicity over sprawling architectures. Chai-1 carried 23 distinct sub-modules; the team deliberately resists adding a "24th module" for a single hard target, favouring simpler, compounding designs that generalise and let it identify clean scaling laws.

Models built from scratch, not fine-tuned LLMs. Everything is built from the ground up on biological data rather than fine-tuning a general language model, drawing on the Protein Data Bank's crystal structures since 1970 plus vast sequence databases and, increasingly, data generated by Chai's own in-house science team.

An Avengers squad of disciplines. Meier came from OpenAI's early team (GPT-1, GPT-2 scaling laws and the ESM protein language models); McPartlon arrived via pure maths and theoretical computer science; later hires added antibody engineers such as Andy Young and product builders from Stripe as each model generation demanded new expertise.

Infrastructure, not its own drug pipeline. Unlike Isomorphic Labs, Chai chose to arm the existing industry rather than develop drugs itself, partnering with Eli Lilly, Novartis, Pfizer and others — a harder, higher-bar business where the models "really have to work," but one the founders argue is more scalable and lets them reinvest in better models.

The vision: a computer-aided design suite for molecules. The goal is not a handful of therapies but changing how medicines are discovered — compressing the idea-to-testable-hypothesis loop from months to weeks or days, and eventually reaching undruggable targets and rare or personalised diseases that are uneconomical today.

Read more: Sequoia Capital · Chai Discovery